M.Y.HUANG1, C.H.QUEK1, Y.H.TING2, W.X.CHEN2, J.SIM3, W.M HUANG4, H.L NG5, P.L.LEUNG1, C.H.TAN2
Singapore Institute of Technology1, Tan Tock Seng Hospital2, MOH Holdings Pte Ltd (MOHH)3, Institute for Infocomm Research4, National Centre for Infectious Diseases5
To determine if machine deep learning using convoluted neural networks (CNN) is able to detect the presence of pulmonary tuberculosis (PTB) changes on chest radiographs (CXRs).
The CXRs acquired from 895 patients seen at our national TB Control Unit (TBCU) over a period from 2010-2019 were utilised for CNN development and testing. CXRs were individually labelled and reviewed by a trained technician and senior radiologist, and separated into four categories: normal, scarring, consolidation and consolidation with cavitation. 19.5% of the labelled CXRs were designated as the validation (test) set. The diagnostic performance of the model was determined through receiver operating characteristics (ROC) analysis.
For determining the presence of cavitation with areas of consolidation (n=244) versus without (n=203), the model performed moderately, with an AUROC of 0.6458. For determining the presence of scarring (n=289) versus without, i.e. normal (n=159), the model performed well, with an AUROC of 0.8122.
Artificial intelligence CNN models show promise for detecting the presence of PTB changes on screening CXRs. For active PTB, the CNN model performs less accurately in detecting the presence of cavitation within areas of consolidation. If corroborated with larger datasets, models of higher performance can be developed to augment review of screening CXRs in the community setting.